Shen, Zhuoyan; (2022) Cancer Outcome Prediction with Multiform Medical Data using Deep Learning. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Cancer_Outcome_Prediction_with_Multiform_Medical_Data_using_Deep_Learning.pdf - Submitted Version Download (35MB) | Preview

Abstract

This thesis illustrated the work done for my PhD project, which aims to develop personalised cancer outcome prediction models using various types of medical data. A predictive modelling workflow that can analyse data with different forms and generate comprehensive outcome prediction was designed and implemented on a variety of datasets. The model development was accompanied by applying deep learning techniques for multivariate survival analysis, medical image analysis and sequential data processing. The modelling workflow was applied to three different tasks: 1. Deep learning models were developed for estimating the progression probability of patients with colorectal cancer after resection and after different lines of chemotherapy, which got significantly better predictive performance than the Cox regression models. Besides, CT-based models were developed for predicting the tumour local response after chemotherapy of patients with lung metastasis, which got an AUC of 0. 769 on disease progression detection and 0.794 on treatment response classification. 2. Deep learning models were developed for predicting the survival state of patients with non-small cell lung cancer after radiotherapy using CT scans, dose distribution and disease and treatment variables. The eventual model obtained by ensemble voting got an AUC of 0.678, which is significantly higher than the score achieved by the radiomics model (0.633). 3. Deep-learning-aided approaches were used for estimating the progression risk for patients with solitary fibrous tumours using digital pathology slides. The deep learning architecture was able to optimise the WHO risk assessment model using automatically identified levels of mitotic activity. Compared to manual counting given by pathologists, deep-learning-aided mitosis counting can re-grade the patients whose risks were underestimated. The applications proved that the predictive models based on hybrid neural networks were able to analyse multiform medical data for generating data-based cancer outcome prediction. The results can be used for realising personalised treatment planning, evaluating treatment quality, and aiding clinical decision-making.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item